CN109858562A

CN109858562A - A kind of classification method of medical image, device and storage medium

Info

Publication number: CN109858562A
Application number: CN201910130011.4A
Authority: CN
Inventors: 高殿宇; 曹世磊; 马锴; 郑冶枫
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2019-02-21
Filing date: 2019-02-21
Publication date: 2019-06-07

Abstract

The embodiment of the invention discloses a kind of classification method of medical image, device and storage mediums；The embodiment of the present invention first obtains multiple medical image samples, classify to the lesion characteristics in the medical image sample, the promotion tree-model of the medical image sample is constructed according to classification results, obtain the first classifier group, filter out the lesion characteristics for meeting the first preset condition from the lesion characteristics using the first classifier group again, obtain target lesion feature set, then, preset second classifier is trained using the target lesion feature set, second classifier after being trained, then, the detection of target lesion feature is carried out to medical image to be sorted based on the second classifier after the training, obtain classification results；The program can effectively improve the accuracy of the classification of medical image.

Description

A kind of classification method of medical image, device and storage medium

Technical field

The present invention relates to fields of communication technology, and in particular to a kind of classification method of medical image, device and storage medium.

Background technique

Liver cancer Microvascular invasion (Microvascular Invasion, MVI) is the important only of influence liver cancer post-operative survival rates Vertical risk factor, but currently, MVI determines after relying on surgical operation acquisition sample that carrying out histopathological examination could be really Fixed, very big limits MVI in the Operative manner and art in clinical practice application to early stage surgical intervention mode, hepatectomy Guidance of adjuvant treatment etc..

Pathologic finding method detects whether that the tumor specimen that MVI depends on surgical operation to obtain, treatment process needs occur Professional was handled sample by the long period, and efficiency is lower, and observed process someone will be subjective judgement, judgement Accuracy rate is lower.

Summary of the invention

The embodiment of the present invention provides classification method, device and the storage medium of a kind of medical image, can effectively improve The accuracy of the classification of medical image.

The embodiment of the present invention provides a kind of classification method of medical image, comprising:

Obtain multiple medical image samples；

Classify to the lesion characteristics in the medical image sample, constructs the medical image according to classification results The promotion tree-model of sample obtains the first classifier group；

The lesion characteristics for meeting the first preset condition are filtered out from the lesion characteristics using the first classifier group, Obtain target lesion feature set；

Preset second classifier is trained using the target lesion feature set, the second classification after being trained Device；

The detection of target lesion feature is carried out to medical image to be sorted based on the second classifier after the training, is divided Class result.

Correspondingly, the embodiment of the present invention also provides a kind of sorter of medical image, comprising:

Acquiring unit, for obtaining multiple medical image samples；

Construction unit is constructed for classifying to the lesion characteristics in the medical image sample according to classification results The promotion tree-model of the medical image sample out obtains the first classifier group；

Screening unit meets the first default item for filtering out from the lesion characteristics using the first classifier group The lesion characteristics of part obtain target lesion feature set；

Training unit is instructed for being trained using the target lesion feature set to preset second classifier Second classifier after white silk；

Detection unit, it is special for carrying out target lesion to medical image to be sorted based on the second classifier after the training Sign detection, obtains classification results.

Optionally, in some embodiments, the construction unit may include extracting subelement and building subelement, such as Under:

The extraction subelement, for being partitioned into lesion region sample in the medical image sample, to the lesion Area sample extracts multidimensional lesion characteristics；

The building subelement constructs described for classifying to the multidimensional lesion characteristics according to classification results The promotion tree-model of medical image sample obtains the first classifier group.

Optionally, in some embodiments, the extraction subelement, specifically for being screened in the medical image sample Tissue regions by lesion region and nearly lesion out；Tissue regions by the lesion region and nearly lesion are extended, are obtained To extension rear region；The extension rear region is split, lesion region sample is obtained；Utilize medical image feature extraction packet To the lesion region sample extraction multidimensional lesion characteristics.

Optionally, in some embodiments, the building subelement is specifically used for using each base classifier from described more Multiple lesion characteristics are randomly choosed in dimension lesion characteristics；The multiple lesion characteristics are classified, it is multiple described to construct The Taxonomy and distribution model of medical image sample；Multiple Taxonomy and distribution models are combined, the medical image sample is obtained Promotion tree-model, obtain the first classifier group.

Optionally, in some embodiments, the multiple medical image sample may include medical image training sample and Medical image verifies sample, and the training unit may include trained subelement and verify subelement, as follows:

The trained subelement, for constructing default second classifier using the target lesion feature set, using described Medical image training sample is trained default second classifier, obtains the second classifier；

The verifying subelement, for being tested using medical image verifying sample the accuracy of second classifier Card, if verification result meets the second preset condition, the second classifier after being trained.

Optionally, in some embodiments, the detection unit may include obtaining subelement, detection sub-unit and determination Subelement, as follows:

The acquisition subelement, for obtaining medical image to be sorted；

The detection sub-unit, for carrying out target lesion to the medical image using the second classifier after the training Feature detection；

The determining subelement, if indicating the medical image for testing result, there are target lesion features, it is determined that There are lesion regions for the medical image.

In addition, the embodiment of the present invention also provides a kind of storage medium, the storage medium is stored with a plurality of instruction, the finger It enables and being loaded suitable for processor, the step in classification method to execute any medical image provided in an embodiment of the present invention Suddenly.

The embodiment of the present invention first obtains multiple medical image samples, divides the lesion characteristics in the medical image sample Class constructs the promotion tree-model of the medical image sample according to classification results, obtains the first classifier group, then using this first Classifier group filters out the lesion characteristics for meeting the first preset condition from the lesion characteristics, obtains target lesion feature set, connects , preset second classifier is trained using the target lesion feature set, the second classifier after being trained, then, The detection of target lesion feature is carried out to medical image to be sorted based on the second classifier after the training, obtains classification results；By Can be first with the feature of medical image sample the first classifier of training in the program, recycling the first classifier to filter out can be just Really reflect the important feature (i.e. target lesion feature) of the lesion, then train the second classifier using the target lesion feature, To improve the accuracy rate of the second classifier classification, it is ensured that the information in medical image can be accurately detected, so, phase For relying solely on manually for the division that medical image presentation information is judged, medical image classification can effectively improve Accuracy.

Detailed description of the invention

To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for For those skilled in the art, without creative efforts, it can also be obtained according to these attached drawings other attached Figure.

Fig. 1 a is the schematic diagram of a scenario of the classification method of medical image provided in an embodiment of the present invention；

Fig. 1 b is the flow chart of the classification method of medical image provided in an embodiment of the present invention；

Fig. 1 c is the structural schematic diagram of the base classifier of medical image provided in an embodiment of the present invention；

Fig. 1 d is the structural schematic diagram of the object classifiers of medical image provided in an embodiment of the present invention；

Fig. 2 a is another flow chart of the classification method of medical image provided in an embodiment of the present invention；

Fig. 2 b is the another flow chart of the classification method of medical image provided in an embodiment of the present invention；

Fig. 2 c is another structural schematic diagram of the base classifier of medical image provided in an embodiment of the present invention；

Fig. 2 d is the another structural schematic diagram of the object classifiers of medical image provided in an embodiment of the present invention；

Fig. 2 e is the schematic diagram of the target lesion feature of medical image provided in an embodiment of the present invention；

Fig. 2 f is the structural schematic diagram of the classification method of medical image provided in an embodiment of the present invention；

Fig. 3 is the structural schematic diagram of the sorter of medical image provided in an embodiment of the present invention；

Fig. 4 is the structural schematic diagram of the network equipment provided in an embodiment of the present invention.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, those skilled in the art's every other implementation obtained without creative efforts Example, shall fall within the protection scope of the present invention.

The embodiment of the present invention provides classification method, device and the storage medium of a kind of medical image.Wherein, the medical image Classification can integrate in the network device, which can be server, be also possible to the equipment such as terminal.

For example, with reference to Fig. 1 a, firstly, the network equipment of the sorter for being integrated with medical image first obtains multiple medicine Image sample classifies to the lesion characteristics in the medical image sample, constructs the medical image sample according to classification results This promotion tree-model obtains the first classifier group, then filters out satisfaction from the lesion characteristics using the first classifier group The lesion characteristics of first preset condition obtain target lesion feature set, then, using the target lesion feature set to preset Two classifiers are trained, the second classifier after being trained, then, based on the second classifier after the training to doctor to be sorted It learns image and carries out the detection of target lesion feature, obtain classification results.

Due to classifying before being detected using the second classifier first with the feature training first of medical image sample Device recycles the first classifier to filter out the important feature (i.e. target lesion feature) that can correctly reflect the lesion, then uses The target lesion feature trains the second classifier, to improve the accuracy rate of the second classifier classification, it is ensured that the letter in medical image Breath can be accurately detected, so, manually divide what medical image presentation information was judged relative to relying solely on For case, the accuracy of medical image classification can effectively improve.

It is described in detail separately below.It should be noted that the following description sequence is not as excellent to embodiment The restriction of choosing sequence.

The angle of sorter from medical image is described the present embodiment, and the sorter of the medical image is specific It can integrate in the network device, which can be server, be also possible to the equipment such as terminal；Wherein, which can To include the equipment such as mobile phone, tablet computer, laptop and individual calculus (Personal Computer, PC).

A kind of classification method of medical image, comprising: multiple medical image samples are first obtained, in the medical image sample Lesion characteristics classify, the promotion tree-model of the medical image sample is constructed according to classification results, obtains the first classification Device group, then the lesion characteristics for meeting the first preset condition are filtered out from the lesion characteristics using the first classifier group, it obtains Target lesion feature set is then trained preset second classifier using the target lesion feature set, after being trained Then second classifier carries out the detection of target lesion feature to medical image to be sorted based on the second classifier after the training, Obtain classification results.

As shown in Figure 1 b, the detailed process of the classification method of the medical image can be such that

101, multiple medical image samples are obtained.

For example, specifically acquiring equipment, such as computed tomographic scanner (Computed by each medical image Tomography, CT) or NMR imaging instrument etc. to carry out image collection to life body tissue, by professional to image It is labeled, for example is marked by image department doctor and then be supplied to the sorter of the medical image, that is, the classification of medical image Device specifically can receive the medical image sample that medical image acquisition equipment is sent.

Wherein, medical image refers in medical treatment or medical research, obtains life entity or life entity with non-intruding mode The image of certain partial interior tissue, such as the images such as brain, stomach, liver, heart, throat and vagina of human body, these images It can be CT images, nuclear magnetic resonance image or the influence of positron emission tomoscan etc..And life entity has referred to life shape Independent individual of state, such as human or animal etc..Medical image sample also refers to acquire via medical image acquisition equipment, The image got by all means, such as from database or network etc..Wherein, which can be via special Industry personnel carry out the image sample of certain sense mark to image, are also possible to the image sample without any processing.

102, classify to the lesion characteristics in the medical image sample, construct the medical image according to classification results The promotion tree-model of sample obtains the first classifier group；For example, specifically can be such that

(1) it is partitioned into lesion region sample in the medical image sample, to the lesion region sample extraction multidimensional lesion Feature.

For example, the tissue regions by lesion region and nearly lesion can be specifically filtered out in the medical image sample, it will Tissue regions by the lesion region and nearly lesion are extended, and be expanded rear region, are split to the extension rear region, Lesion region sample is obtained, using medical image feature extraction packet to the lesion region sample extraction multidimensional lesion characteristics, is utilized The multidimensional lesion characteristics construct the training set of the first classifier group.

For example, medical image sample is to be labelled with liver cancer data by image department doctor using medical image as Hepatic CT figure For Hepatic CT figure, liver cancer diseased region and nearly Para-cancerous tissue area can be specifically screened in the region comprising liver cancer of mark Domain, i.e. Microvascular invasion region that may be present, that is, area-of-interest (Region of Interest, ROI), are being screened Region out, which extends outwardly, carries out multiple pixel-expansions, and be expanded rear region, is split, obtains micro- to the extension rear region Vascular invasion area sample, by the python packet (open source packet pyradiomics) of the open source of medical image feature extraction from multi-party Face (such as from image modality, first order statistic, shape information, gray level co-occurrence matrixes, gray scale run-length matrix, gray areas size Matrix, two dimensional gray difference matrix, gray level correlative matrix etc.) it is automatically performed and each Microvascular invasion area sample is extracted The feature of multidimensional reaction MVI information.

Wherein, image modality can from raw video, wavelet filtering, Laplacian filtering, square, it is square root, right Several and index angularly extracts, and first order statistic can be from energy (Energy), entropy (Entropy), minimum value (Minimum), the 10th percentage point (10th percentile), the 90th percentage point of (90th percentile), maximum value (Maximum), average value (Mean) and kurtosis (Kurtosis) are angularly extracted, and shape information can be from mesh volume (Mesh Volume), voxel volume (Voxel Volume), surface area (Surface Area), spherical (Sphericity), most Big three-dimensional diameter (Maximum3D diameter), main axis length (Major Axis Length) and flatness (Flatness) it angularly extracts, gray level co-occurrence matrixes can be from auto-correlation (Autocorrelation), the average (Joint of joint Average), Clustering Tendency (Cluster Tendency), contrast (Contrast), correlation (Correlation), difference Entropy (Difference Entropy) and maximum correlation coefficient (Maximal Correlation Coefficient) isogonism Degree extracts, and gray scale run-length matrix can be from the short distance of swimming factor (Short Run Emphasis), the long distance of swimming factor (Long Run Emphasis), gray scale inhomogeneities (Gray Level Non-Uniformity), distance of swimming sum percentage (Run Percentage), gray variance (Gray Level Variance) and the low ash degree distance of swimming factor (Low Gray Level Run Emphasis) angularly extract, the big minor matrix of gray areas can be from the zonule factor (Small Area Emphasis), big Location factor (Large Area Emphasis), area percentage (Zone Percentage), gray scale side Poor (Gray Level Variance) and the low gray level areas factor (Low Gray Level Zone Emphasis) isogonism Degree extracts, and two dimensional gray difference matrix can be from roughening (Coarseness), contrast (Contrast), busy degree (Busyness), complexity (Complexity) and intensity (Strength) are angularly extracted, and gray level correlative matrix can be from Small dependent factor (Small Dependence Emphasis), big dependent factor (Large Dependence Emphasis), And it relies on variance (Dependence Variance) and angularly extracts.

(2) classify to the multidimensional lesion characteristics, the boosted tree of the medical image sample is constructed according to classification results Model obtains the first classifier group.

For example, the first classifier group includes multiple base classifiers, each base classifier specifically can use from multidimensional disease Become in feature and randomly choose multiple lesion characteristics, multiple lesion characteristics is classified, to construct multiple medical images The Taxonomy and distribution model of sample combines multiple Taxonomy and distribution models, constructs the boosted tree mould of the medical image sample Type obtains the first classifier group.

Wherein, Taxonomy and distribution (Classification And Regression Tree, CART) is decision tree One kind, and it is very important decision tree, it is a binary tree, and there are two children for each non-leaf nodes, so right In the first stalk tree its leaf node number 1 more than non-leaf nodes number.Decision tree is a kind of one for relying on decision and setting up Kind tree.In machine learning, decision tree is a kind of prediction model, and representative is a kind of one kind between object properties and object value Mapping relations, some object of each node on behalf, each of tree diverging paths represent some possible attribute value, and every One leaf node then corresponds to the value of object represented by from root node to leaf node path experienced.Classification and recurrence Tree is the learning method of the conditional probability distribution of output stochastic variable Y under the conditions of given input stochastic variable X.Classification and recurrence Tree assumes that decision tree is binary tree, and the value of internal node feature is "Yes" and "No", and left branch is the branch that value is "Yes", Right branch is the branch that value is "No".Such decision tree is equivalent to recursively two points of each features, is spy by the input space Sign space is divided into limited unit, and the probability distribution of prediction is determined on these units, that is, the item given in input The conditional probability distribution exported under part.Taxonomy and distribution algorithm mainly includes the generation of tree and the beta pruning of tree.It is primarily based on instruction Practice data set generation decision tree, the decision tree of generation is big as far as possible；Then beta pruning is carried out to generated tree with validation data set And optimal subtree is selected, at this moment standard of the loss function minimum as beta pruning.The generation of decision tree is exactly to pass through recursively to construct The process of binary decision tree, minimizes criterion to regression tree square error, minimizes criterion to classification tree gini index, into Row feature selecting generates binary tree.

Wherein, it promotes tree-model and uses addition model (linear combination of basic function) and forward direction substep algorithm, while base letter Number uses decision Tree algorithms, treats classification problem using binary class tree, uses y-bend regression tree for regression problem.Boosted tree Model can be regarded as the addition model of decision tree, first train a base classifier from initial training collection, classify further according to base The performance of device is adjusted training sample distribution, so that the training sample that previously base classifier had done wrong is subsequent by more Concern, is then based on sample distribution adjusted to train next base classifier；So repeat, until base classifier number Mesh reaches value T specified in advance, this T base learner is finally weighted combination, get a promotion tree-model.It namely will be weak Classifier is converted into object classifiers, i.e., the very low classifier of many classification accuracies by updating the weight to data, collection At the classifier to form a good classification effect that gets up.

For example, each base classification implement body can be the core using Taxonomy and distribution model as algorithm, such as Fig. 1 c institute Show, for sample set D { sample 1, sample 2 ... sample i ... sample n }, wherein sample 1 includes t1, t2 ... tm, sample i The tm including t1, t2 ..., sample n include t1, t2 ... tm.Wherein, each sample characteristics include a variety of values.Classification and recurrence The building process of tree is as follows:

Firstly, initializing to all samples in sample set D, then, the root node of a Taxonomy and distribution is generated D, and sample set D is distributed to and makees root node d.

Sample set D is divided into two sub- sample set D1 { samples 1, sample 2 ... sample based on feature "Yes" or "No" is divided This k } and D2 { sample k+1 ... sample n }；Then, two child nodes d1 and d2 for generating present node d, distribute D1 to height D2 is distributed to child node d2 by node d1.

Then, for each child node, by taking child node d1 as an example, judge whether child node meets default classification and terminate item Part, if so, using current child node d1 as leaf node, and according to the class of the corresponding subsample concentration sample of child node d1 Leaf node output is not set.

When child node is unsatisfactory for default classification termination condition, continue to classify to the corresponding subsample collection of child node, Such as by taking child node d1 as an example, D1 is divided into two sub- sample sets based on feature "Yes" or "No" and division points are divided, it can D1 is divided into subsample collection D11, D12；Then, it generates child node d11, d12 of present node d1, distinguish D11, D12 Child node d11, d12 is distributed to, for another example by taking child node d2 as an example, is based on dividing feature "Yes" or "No" and division points for D2 Two sub- sample sets are divided into, D2 can be divided into subsample collection D21, D22；Then, the child node of present node d2 is generated D21, d22, D21, D22 are respectively allocated to child node d21, d22.And so on, it is terminated until child node meets default classification Condition.Taxonomy and distribution model in the multiple base classifiers of circuit training finally combines multiple base classifiers and gets a promotion tree Model.

Wherein, algorithm specifically can be such that

Sample set D is constituted by the medical image sample got, the input sample collection D and loss function L in base classifier, Such as:

Loss function is L (y, f (x)),

Wherein, i is characterized ordinal number, x⁽ⁱ⁾For some feature of sample, n is the characteristic that single sample extracts, and m is to choose The characteristic selected, RⁿFor the multidimensional lesion characteristics data set of total sample, X is training sample, and x is the sample for inputting classifier, y It is characterized threshold value, f (x) is classification results of the classifier to sample.

To Taxonomy and distribution model initialization, estimation is that the constant value of loss function minimization (is an only root section The tree of point), i.e.,

Wherein, f₀(x) it refers to first node initializing.

Then, K CART base classifier of circuit training, wherein k=1,2 ..., K, such as:

(a) value of the negative gradient in "current" model of loss function is calculated, using it as the estimation of residual error, such as: for i=1, 2 ..., m

Wherein, r_kiRefer to the residual error of some feature.

(b) node region for estimating regression tree, with the approximation of regression criterion, such as regression criterion r_ki, learn a subtree, Obtain the leaf node region R of kth tree_kj, wherein node j=1,2 ..., J, R_kjFor the residual sum of all features, i.e. r_ki's With.

(c) using the value in linear search estimation leaf node region, make loss function minimization, to j=1,2 ..., J, meter It calculates:

Wherein, c_kjFor the minimum value of residual error, C is constant.

(d) regression tree is updated:

Wherein, I is the value of internal node feature.

Then, multiple CART base classifiers are combined, get a promotion tree-modelIt is as follows:

103, the lesion characteristics for meeting the first preset condition are filtered out from the lesion characteristics using the first classifier group, Obtain target lesion feature set.

Wherein, the setting means of preset condition can there are many kinds of, for example, can flexibly be set according to the demand of practical application It sets, storage can also be pre-set in the network device.In addition, preset condition can be built in the network equipment, alternatively, Can save in memory and be sent to the network equipment, etc..Wherein, preset condition can be in the highest base of accuracy rate point The highest feature of the frequency of occurrences in class device is also possible to biggish feature of importance value, etc. in base classifier.Wherein, Target lesion feature refers to the important feature that can most reflect lesion self character, and target lesion feature set refers to multiple targets The set of lesion characteristics.

For example, specifically cross validation can be carried out to multiple base classifiers in the first classifier group, each base is obtained The accuracy rate of classifier, the multiple base classifiers for selecting the accuracy rate to be greater than preset threshold are combined, and obtain object classifiers, As shown in Figure 1 d, target lesion feature set is filtered out based on the object classifiers.

Wherein, cross validation (Cross Validation), also referred to as circulation estimation (Rotation Estimation), just It is to be grouped initial data under certain meaning, a part is used as training set, and another part is as verifying collection.First with instruction Practice collection to be trained classifier, recycles verifying collection to test the model that training obtains, with this as classification of assessment device Performance indicator.

Wherein, the setting means of preset threshold can there are many kinds of, for example, can flexibly be set according to the demand of practical application It sets, storage can also be pre-set in the network device.In addition, preset threshold can be built in the network equipment, alternatively, Can save in memory and be sent to the network equipment, etc..

Wherein, object classifiers refer to that multiple accuracys rate are greater than the combination of the base classifier of preset threshold.

Optionally, in order to more accurately filter out target lesion feature, can first count base classifier accuracy rate it is higher when Used lesion characteristics screen to obtain target lesion feature according to frequency of use, i.e. step " is screened based on the object classifiers Target lesion feature set out ", comprising:

It counts base classifier accuracy rate in the cross validation and is greater than used lesion characteristics when preset threshold, obtain lesion The frequency of use of feature；

The highest multiple lesion characteristics of frequency of use are filtered out according to statistical result, obtain target lesion feature set.

104, preset second classifier is trained using the target lesion feature set, the second classification after being trained Device.

For example, multiple medical image sample can be specifically divided into medical image training sample and medical image verifying sample This, constructs default second classifier first with the target lesion feature set, default to this using the medical image training sample Second classifier is trained, and obtains the second classifier, then, using medical image verifying sample to the standard of second classifier True property is verified, if verification result meets the second preset condition, the second classifier after being trained.

For example, default second classification of important feature (the i.e. target lesion feature) building screened specifically can be used Device excludes the negative effect that insignificant feature generates the building of the second classifier, then adjusts the parameter of classifier, Optimum Classification Device generates final mask (i.e. the second classifier).Since the second classifier uses the important feature screened to carry out structure It builds, train and arameter optimization, then carry out test and eliminating the interference that other insignificant features constructs classifier, it is available Better result (i.e. more optimized classifier).

105, the detection of target lesion feature is carried out to medical image to be sorted based on the second classifier after the training, obtained Classification results.

For example, specifically target lesion feature can be carried out to medical image to be sorted based on the second classifier after the training Detection, according to testing result classifies to medical image to be sorted, obtains classification results.

For example, specific available medical image to be sorted；Using the second classifier after the training to the medical image Carry out the detection of target lesion feature；If testing result indicates the medical image, there are target lesion features, it is determined that the medicine shadow As there are lesion regions.

Wherein, medical image to be sorted refers to the medical image classified, and can be adopted by each medical image Collect equipment, for example computed tomographic scanner or NMR imaging instrument etc. to carry out image collection to life body tissue, Such as brain, stomach, liver, heart, throat and vagina of human body etc., and then it is supplied to the medical image detection device, that is, doctor Learning image detection device specifically can receive the medical image to be sorted that medical image acquisition equipment is sent.

From the foregoing, it will be observed that the embodiment of the present invention first obtains multiple medical image samples, to the lesion in the medical image sample Feature is classified, and the promotion tree-model of the medical image sample is constructed according to classification results, obtains the first classifier group, then The lesion characteristics for meeting the first preset condition are filtered out from the lesion characteristics using the first classifier group, obtain target lesion Feature set is then trained preset second classifier using the target lesion feature set, the second classification after being trained Then device carries out the detection of target lesion feature to medical image to be sorted based on the second classifier after the training, is classified As a result；Since the program can recycle the first classifier sieve first with the feature of medical image sample the first classifier of training The important feature (i.e. target lesion feature) that can correctly reflect the lesion is selected, then using target lesion feature training second Classifier, to improve the accuracy rate of the second classifier classification, it is ensured that the information in medical image can be accurately detected, So can effectively improve medicine relative to relying solely on manually for the division that medical image presentation information is judged The accuracy of image classification.

According to method described in upper one embodiment, citing is described in further detail below.

In the present embodiment, it will be specifically integrated in the network equipment with the sorter of the medical image, with medical image Specially Hepatic CT figure be illustrated for the detection of hepatocellular carcinoma Microvascular invasion to Hepatic CT figure.

As shown in Figure 2 a, a kind of classification method of medical image, detailed process can be such that

201, the network equipment obtains multiple medical image samples.

For example, the network equipment can receive the CT scan image of user's input, set alternatively, receiving other The CT scan image that preparation is sent, wherein the CT scan image can be calculated by electronics Machine tomography scanner carries out image collection to liver area, and the image got by all means, for example, from database or Then network etc. is marked by image department doctor, obtain multiple liver CT images samples, and then be supplied to the network equipment.

202, the network equipment is partitioned into lesion region sample in the medical image sample, to the lesion region sample extraction Multidimensional lesion characteristics.

For example, the network equipment can specifically screen liver by liver cancer diseased region and nearly cancer in the region comprising liver cancer of mark Tissue regions, as shown in Figure 2 b, i.e. Microvascular invasion region that may be present, that is, area-of-interest, in the area filtered out Domain, which extends outwardly, carries out multiple pixel-expansions, and be expanded rear region, is split to the extension rear region, obtains capilary and invade Violate area sample, by the open source of medical image feature extraction python packet (open source wrap pyradiomics) from many aspects (such as From image modality, first order statistic, shape information, gray level co-occurrence matrixes, gray scale run-length matrix, the big minor matrix of gray areas, two Dimension grey scale difference matrix, gray level correlative matrix etc.) it is automatically performed that extract multidimensional to each Microvascular invasion area sample anti- Answer the feature of MVI information.

For example, image modality can from raw video, wavelet filtering, Laplacian filtering, square, it is square root, right Several and index angularly extracts, and first order statistic can be from energy, entropy, minimum value, the 10th percentage point, the 90th percentage Point, maximum value, average value and kurtosis are angularly extracted, and shape information can be from mesh volume, voxel volume, surface area, ball Shape, maximum three-dimensional diameter, main axis length and flatness are angularly extracted, and gray level co-occurrence matrixes can be flat from auto-correlation, joint , Clustering Tendency, contrast, correlation, Difference Entropy and maximum correlation coefficient angularly extract, and gray scale run-length matrix can be with From the short distance of swimming factor, the long distance of swimming factor, gray scale inhomogeneities, distance of swimming sum percentage, gray variance and the low ash degree distance of swimming because Son angularly extract, the big minor matrix of gray areas can from the zonule factor, big Location factor, area percentage, gray variance, And the low gray level areas factor is angularly extracted, two dimensional gray difference matrix can be from roughening, contrast, busy degree, complexity Degree and intensity are angularly extracted, and gray level correlative matrix can be from small dependent factor, big dependent factor and dependence variance etc. Angle extraction extracts 1217 features for reacting the MVI information to each Microvascular invasion area sample.

203, the network equipment classifies to the multidimensional lesion characteristics, constructs the medical image sample according to classification results Promotion tree-model, obtain the first classifier group.

For example, the extreme gradient promotion of network equipment calling python open source packet (eXtreme Gradient Boosting, XGBoost the XGBoost classifier in) sets the number and depth set in classifier, minimum sample in each leaf node Then the parameters such as number, learning rate, number of training accounting, feature accounting and costing bio disturbance function utilize each base classifier Multiple lesion characteristics are randomly choosed from the multidimensional lesion characteristics, for example the feature selecting 974 of the MVI information is reacted from 1217 A feature is classified by 974 lesion characteristics inputs and into classifier, to construct multiple medical image samples Taxonomy and distribution model calls the GridSearchCV in scikit-learn to carry out arameter optimization, improves the standard of classifier True rate.Multiple Taxonomy and distribution models are combined, the promotion tree-model of the medical image sample is obtained.

Wherein, scikit-learn is the machine learning algorithm library realized with python.Scikit-learn may be implemented The common machine learning algorithms such as data prediction, classification, recurrence, dimensionality reduction, model selection.And GridSearchCV is exactly automatic Ginseng is adjusted, as long as parameter is input into, the result and parameter of optimization can be provided.

For example, each base classification implement body can be the core using Taxonomy and distribution model as algorithm, for sample Collect D { sample 1, sample 2 ... sample i ... sample n }, wherein sample 1 include t1, t2 ... tm, sample i include t1, T2 ... tm, sample n include t1, t2 ... tm.Wherein, each sample characteristics include a variety of values.The structure of Taxonomy and distribution It is as follows to build process:

Then, for each child node, by taking child node d1 as an example, judge whether child node meets default classification and terminate item Part, for example, specifically, as shown in Figure 2 c, presetting classification termination condition can be leaf=0.0019, if so, by current Child node d1 according to the corresponding subsample child node d1 concentrates the classification of sample that the leaf node is arranged is defeated as leaf node Out.

Wherein, algorithm specifically can be such that

Loss function is L (y, f (x)),

Wherein, f₀(x) it refers to first node initializing.

Wherein, r_kiRefer to the residual error of some feature.

Wherein, c_kjFor the minimum value of residual error, C is constant.

(d) regression tree is updated:

Wherein, I is the value of internal node feature.

For example, specifically, as shown in Figure 2 c, firstly, initialization carried out to all samples in sample set D choose initial value be f₀Then < 90.2232 generates the root node d an of Taxonomy and distribution, and sample set D is distributed to and makees root node d.

Then, for each child node, judge whether child node meets default classification termination condition leaf=0.0019, by Child node d1 known to figure meets default classification termination condition, then using current child node d1 as leaf node, and according to sub- section Point d1 concentrates corresponding subsample the classification of sample that leaf node output is arranged.

Child node d2 is f₀When < 571.7720 is unsatisfactory for default classification termination condition, continue to the corresponding son of child node Sample set is classified, and D2 is divided into two sub- sample sets based on feature "Yes" or "No" and division points are divided, can be incited somebody to action D2 is divided into subsample collection D21, D22；Then, it generates child node d21, d22 of present node d2, distribute D21, D22 respectively To child node d21, d22, the value for choosing d21 is f₀< 245.9164, the value for choosing d22 are f₀< 4336.0493, to child node D21, d22 continue to divide.And so on, until child node meets default classification termination condition.The multiple base classifiers of circuit training In Taxonomy and distribution model, finally combine multiple base classifiers and get a promotion tree-model.

204, the network equipment is filtered out from the lesion characteristics using the first classifier group meets the first preset condition Lesion characteristics obtain target lesion feature set.

Wherein, the setting means of preset condition can there are many kinds of, for example, can flexibly be set according to the demand of practical application It sets, storage can also be pre-set in the network device.In addition, preset condition can be built in the network equipment, alternatively, Can save in memory and be sent to the network equipment, etc..Wherein, preset condition can be in the highest base of accuracy rate point The highest feature of the frequency of occurrences in class device is also possible to biggish feature of importance value, etc. in base classifier.

For example, the first classifier group includes multiple base classifiers, the network equipment can be pressed all medical image samples It is training sample and verifying sample according to 4:1 points, multiple base classifiers in the first classifier group can be specifically carried out 500 times 5 folding cross validations obtain the accuracy rate of each base classifier, then, select the accuracy rate to be greater than preset threshold, for example, accurately Multiple base classifiers of the rate greater than 90% are combined, and obtain object classifiers, for example, as shown in Figure 2 d, based on the target point Class device filters out target lesion feature set, for example, as shown in Figure 2 e, filter out 20 important features (i.e. target lesion feature, Be exactly the feature sensitive to MVI) and its numerical value of importance indicate that removing other insignificant features influences, for example, filtering out 20 target lesion features can be such that

1 is original _ shape _ elongation (original_shape_Elongation), and 2 be logarithmic function-standard deviation -2- 0- millimeters _ three-dimensional _ single order _ minimum value (log-sigma-2-0-mm-3D_firstorder_Minimum), 3 be logarithmic function- - 5-0- millimeters of standard deviation _ three-dimensional _ gray scale and with the matrix _ big area grayscale factor (log-sigma-5-0-mm-3D_glszm_ LargeAreaLowGrayLevelEmphsis), 4 be small echo _ low low-and high-frequency subband _ gray level co-occurrence matrixes _ minimum material requirement 1 (wavelet_LHL_glcm_lmc1), 5 be small echo _ low high high-frequency sub-band _ gray level co-occurrence matrixes _ cluster shade (wavelet_ LHH_glcm_ClusterShade), 6 be small echo _ low low frequency sub-band _ single order _ intermediate value (wavelet_LLL_firstorder_ Median), 7 it is sliced (original_shape_Maximum2DDiameterSlice) for the two-dimentional diameter of original _ shape _ maximum, 8 be logarithmic function--5-0- millimeters of standard deviation _ three-dimensional _ single order _ 10th percentage point (log-sigma-5-0-mm-3D_ Firstorder_10Percentile), 9 be logarithmic function--1-0- millimeters of standard deviation _ three-dimensional _ based on gray level co-occurrence matrixes Texture feature extraction _ dependence inhomogeneities standardizes (log-sigma-1-0-mm-3D_gldm_DependenceNonUinfor MityNormalized), 10 be logarithmic function--3-0- millimeters of standard deviation _ three-dimensional _ gray level co-occurrence matrixes _ inverse variance (log- Sigma-3-0-mm-3D_glcm_InverseVariance), 11 be logarithmic function--5-0- millimeters of standard deviation _ three-dimensional _ gray scale Stroke matrix _ long the distance of swimming factor (log-sigma-5-0-mm-3D_glrlm_LongRunEmphasis), 12 be small echo _ low height High-frequency sub-band _ gray level co-occurrence matrixes _ Clustering Tendency (wavelet_LHH_glcm_ClusterTendency), 13 be small echo _ height Low-and high-frequency subband _ gray level co-occurrence matrixes _ correlation (wavelet_HHL_glcm_Correlation), 14 be small echo _ Gao Gaogao Frequency subband _ single order _ intermediate value (wavelet_HHH_firstorder_Median), 15 be original _ single order _ intermediate value (original_ Firstorder_Median), 16 be logarithmic function--4-0- millimeters of standard deviation _ big minor matrix of three-dimensional _ gray areas _ big region The low ash degree factor (log-sigma-4-0-mm-3D_glszm_LargeAreaLowGrayLevelEmphsis), 17 for small echo _ Low low-and high-frequency subband _ single order _ interquartile range (wavelet_LHL_firstorder_InterquartileRange), 18 be small Wave _ low high high-frequency sub-band _ single order _ intermediate value (wavelet_LHH_firstorder_Median), 19 be original _ gray scale stroke square Battle array _ long the distance of swimming factor (original_glrlm_LongRunEmphasis), 20 be original _ texture based on gray level co-occurrence matrixes Feature extraction _ dependence variance (original_gldm_DependenceVariance).

205, the network equipment is trained preset second classifier using the target lesion feature set, after being trained Second classifier.

For example, it is second point default that 20 important features (the i.e. target lesion feature) building screened specifically can be used Class device excludes the negative effect that insignificant feature generates the building of the second classifier, then adjusts the parameter of classifier, optimization point Class device generates final mask (i.e. the second classifier).Since the second classifier uses the important feature screened to carry out Building, trained and arameter optimization, are then tested, eliminate the interference that other insignificant features construct classifier, can Obtain better result (i.e. more optimized classifier).If verification result meets the second preset condition, for example, the second classifier Accuracy rate is greater than 90% etc., then the second classifier after being trained.

206, the network equipment carries out target lesion feature to medical image to be sorted based on the second classifier after the training Detection, obtains classification results.

For example, the specific available medical image to be sorted of the network equipment, using the second classifier after the training to this Medical image carries out the detection of target lesion feature, if testing result indicates the medical image, there are target lesion features, it is determined that There are lesion regions for the medical image.

For example, as shown in figure 2f, the front-end A of the network equipment receives data (hepatocellular carcinoma CT imaging inputted), then on It is transmitted to rear end, rear end is determined using the second classifier after training, is being output to front end B.

In addition, it should be noted that, the execution hardware environment of the program can according to actual needs depending on.

In addition, it should be noted that, the present embodiment is only example, it should be appreciated that is used in the present embodiment Classifier can also use other classifiers to carry out feature selecting and construct to divide other than it can be Taxonomy and distribution model Class device, etc..

From the foregoing, it will be observed that the network equipment of the embodiment of the present invention can first obtain multiple medical image samples, to the medical image Lesion characteristics in sample are classified, and the promotion tree-model of the medical image sample is constructed according to classification results, obtain One classifier group, then the lesion spy for meeting the first preset condition is filtered out from the lesion characteristics using the first classifier group Sign, is obtained target lesion feature set, then, is trained, is obtained to preset second classifier using the target lesion feature set Then second classifier after to training carries out target lesion to medical image to be sorted based on the second classifier after the training Feature detection, obtains classification results；Since the program can train the first classifier first with the feature of medical image sample, then The important feature (i.e. target lesion feature) that can correctly reflect the lesion is filtered out using the first classifier, then uses the target Lesion characteristics train the second classifier, to improve the accuracy rate of the second classifier classification, it is ensured that information in medical image can be with It is accurately detected；And CT images in the preoperative are utilized, a large amount of of reaction tumor information are automatically extracted by area of computer aided Feature, accurately determines whether the life entity has MVI.Entire decision process is all that computer is automatically performed, quickly and effectively, Without subjective factor, the confidence level of algorithm is increased, and feature selection process directly extracts the important feature of high-accuracy classifier, Feature i.e. sensitive to MVI, is added without subjective judgement, increases the robustness of model.Therefore, the present invention can be improved doctor and exist The preoperative understanding to hepatocarcinoma patient MVI situation occurred, in patient's early stage surgical intervention mode, the Operative manner of hepatectomy, art Adjuvant treatment etc. is instructed, so, manually divide what medical image presentation information was judged relative to relying solely on For case, the accuracy of medical image classification can effectively improve, promotion judges effect, improves prognosis.

In order to better implement above method, correspondingly, the embodiment of the present invention also provides a kind of classification dress of medical image It sets, the sorter of the medical image specifically can integrate in the network device, which can be server, can also be with It is the equipment such as terminal.

For example, as shown in figure 3, the sorter of the medical image may include acquiring unit 301, construction unit 302, sieve Menu member 303, training unit 304 and detection unit 305 are as follows:

(1) acquiring unit 301；

Acquiring unit 301, for obtaining multiple medical image samples.

For example, specifically acquiring equipment, such as computed tomographic scanner or Magnetic resonance imaging by each medical image Instrument etc. to carry out image collection to life body tissue, is labeled by professional to image, for example marked by image department doctor And then it is supplied to the acquiring unit 301, that is, the acquiring unit 301 specifically can receive the doctor that medical image acquisition equipment is sent Learn image sample.

(2) construction unit 302；

Construction unit 302 is constructed for classifying to the lesion characteristics in the medical image sample according to classification results The promotion tree-model of the medical image sample out obtains the first classifier group.

Optionally, in some embodiments, which may include extracting subelement and building subelement, such as Under:

Subelement is extracted, for being partitioned into lesion region sample in the medical image sample, to the lesion region sample Extract multidimensional lesion characteristics.

For example, subelement is extracted, specifically for filtering out by lesion region and nearly lesion in the medical image sample Tissue regions by the lesion region and nearly lesion are extended by tissue regions, and be expanded rear region, to the extension back zone Domain is split, and obtains lesion region sample, using medical image feature extraction packet to the lesion region sample extraction from multi-party Face (such as from image modality, first order statistic, shape information, gray level co-occurrence matrixes, gray scale run-length matrix, gray areas size Matrix, two dimensional gray difference matrix, gray level correlative matrix etc.) multidimensional lesion characteristics.

Wherein, image modality can from raw video, wavelet filtering, Laplacian filtering, square, it is square root, right Several and index angularly extracts, and first order statistic can be from energy, entropy, minimum value, the 10th percentage point, the 90th percentage Point, maximum value, average value and kurtosis are angularly extracted, and shape information can be from mesh volume, voxel volume, surface area, ball Shape, maximum three-dimensional diameter, main axis length and flatness are angularly extracted, and gray level co-occurrence matrixes can be flat from auto-correlation, joint , Clustering Tendency, contrast, correlation, Difference Entropy and maximum correlation coefficient angularly extract, and gray scale run-length matrix can be with From the short distance of swimming factor, the long distance of swimming factor, gray scale inhomogeneities, distance of swimming sum percentage, gray variance and the low ash degree distance of swimming because Son angularly extract, the big minor matrix of gray areas can from the zonule factor, big Location factor, area percentage, gray variance, And the low gray level areas factor is angularly extracted, two dimensional gray difference matrix can be from roughening, contrast, busy degree, complexity Degree and intensity are angularly extracted, and gray level correlative matrix can be from small dependent factor, big dependent factor and dependence variance etc. Angle extraction.

It constructs subelement and constructs the medical image according to classification results for classifying to the multidimensional lesion characteristics The promotion tree-model of sample obtains the first classifier group.

Optionally, the first classifier group includes multiple base classifiers, which is specifically used for utilizing each base point Class device randomly chooses multiple lesion characteristics from the training set, and multiple lesion characteristics are classified, to construct multiple be somebody's turn to do The Taxonomy and distribution model of medical image sample combines multiple Taxonomy and distribution models, constructs the medical image sample Tree-model is promoted, the first classifier group is obtained.

(3) screening unit 303；

Screening unit 303 meets the first default item for filtering out from the lesion characteristics using the first classifier group The lesion characteristics of part obtain target lesion feature set.

For example, screening unit 303 specifically can carry out cross validation to multiple base classifiers in the first classifier group, The accuracy rate of each base classifier is obtained, then, the multiple base classifiers for selecting the accuracy rate to be greater than preset threshold are combined, Object classifiers are obtained, filter out target lesion feature set based on the object classifiers.

Optionally, in order to more accurately filter out target lesion feature, can first count base classifier accuracy rate it is higher when Used lesion characteristics screen to obtain target lesion feature according to frequency of use, it may be assumed that

It counts base classifier accuracy rate in the cross validation and is greater than used lesion characteristics when preset threshold, obtain lesion The frequency of use of feature filters out the highest multiple lesion characteristics of frequency of use according to statistical result, obtains target lesion feature Collection.

(4) training unit 304；

Training unit 304 is instructed for being trained using the target lesion feature set to preset second classifier Second classifier after white silk.

Optionally, in some embodiments, multiple medical image sample may include medical image training sample and doctor It learns image and verifies sample, which may include trained subelement and verifies subelement, as follows:

Training subelement, for constructing default second classifier using the target lesion feature set, using the medical image Training sample is preset the second classifier to this and is trained, and the second classifier is obtained；

Subelement is verified, for being verified using medical image verifying sample to the accuracy of second classifier, if Verification result meets the second preset condition, then the second classifier after being trained.

(5) detection unit 305

Detection unit 305, for carrying out target lesion to medical image to be sorted based on the second classifier after the training Feature detection, obtains classification results.For example, specifically can based on the second classifier after the training to medical image to be sorted into The detection of row target lesion feature, according to testing result classifies to medical image to be sorted, obtains classification results.

Optionally, in some embodiments, which may include obtaining subelement, detection sub-unit and determining sub Unit, as follows:

Subelement is obtained, for obtaining medical image to be sorted；

Detection sub-unit, for carrying out the inspection of target lesion feature to the medical image using the second classifier after the training It surveys；

Determine subelement, there are target lesion features if indicating the medical image for testing result, it is determined that the medicine There are lesion regions for image.

When it is implemented, above each unit can be used as independent entity to realize, any combination can also be carried out, is made It is realized for same or several entities, the specific implementation of above each unit can be found in the embodiment of the method for front, herein not It repeats again.

From the foregoing, it will be observed that the embodiment of the present invention first obtains multiple medical image samples by acquiring unit 301, by construction unit Lesion characteristics in 302 pairs of medical image samples are classified, and construct mentioning for the medical image sample according to classification results Tree-model is risen, obtains the first classifier group, then screened from the lesion characteristics using the first classifier group by screening unit 303 The lesion characteristics for meeting the first preset condition out obtain target lesion feature set, then, utilize the target by training unit 304 Lesion characteristics collection is trained preset second classifier, the second classifier after being trained, then, by detection unit 305 The detection of target lesion feature is carried out to medical image to be sorted based on the second classifier after the training, obtains classification results；By Can be first with the feature of medical image sample the first classifier of training in the program, recycling the first classifier to filter out can be just Really reflect the important feature (i.e. target lesion feature) of the lesion, then train the second classifier using the target lesion feature, To improve the accuracy rate of the second classifier classification, it is ensured that the information in medical image can be accurately detected, so, phase For relying solely on manually for the division that medical image presentation information is judged, medical image classification can effectively improve Accuracy.

In addition, the embodiment of the present invention also provides a kind of network equipment, as shown in figure 4, it illustrates institutes of the embodiment of the present invention The structural schematic diagram for the network equipment being related to, specifically:

The network equipment may include one or more than one processing core processor 401, one or more The components such as memory 402, power supply 403 and the input unit 404 of computer readable storage medium.Those skilled in the art can manage It solves, network equipment infrastructure shown in Fig. 4 does not constitute the restriction to the network equipment, may include more more or fewer than illustrating Component perhaps combines certain components or different component layouts.Wherein:

Processor 401 is the control centre of the network equipment, utilizes various interfaces and connection whole network equipment Various pieces by running or execute the software program and/or module that are stored in memory 402, and are called and are stored in Data in reservoir 402 execute the various functions and processing data of the network equipment, to carry out integral monitoring to the network equipment. Optionally, processor 401 may include one or more processing cores；Preferably, processor 401 can integrate application processor and tune Demodulation processor processed, wherein the main processing operation system of application processor, user interface and application program etc., modulatedemodulate is mediated Reason device mainly handles wireless communication.It is understood that above-mentioned modem processor can not also be integrated into processor 401 In.

Memory 402 can be used for storing software program and module, and processor 401 is stored in memory 402 by operation Software program and module, thereby executing various function application and data processing.Memory 402 can mainly include storage journey Sequence area and storage data area, wherein storing program area can the (ratio of application program needed for storage program area, at least one function Such as sound-playing function, image player function) etc.；Storage data area, which can be stored, uses created number according to the network equipment According to etc..In addition, memory 402 may include high-speed random access memory, it can also include nonvolatile memory, such as extremely A few disk memory, flush memory device or other volatile solid-state parts.Correspondingly, memory 402 can also wrap Memory Controller is included, to provide access of the processor 401 to memory 402.

The network equipment further includes the power supply 403 powered to all parts, it is preferred that power supply 403 can pass through power management System and processor 401 are logically contiguous, to realize management charging, electric discharge and power managed etc. by power-supply management system Function.Power supply 403 can also include one or more direct current or AC power source, recharging system, power failure monitor The random components such as circuit, power adapter or inverter, power supply status indicator.

The network equipment may also include input unit 404, which can be used for receiving the number or character of input Information, and generate keyboard related with user setting and function control, mouse, operating stick, optics or trackball signal Input.

Although being not shown, the network equipment can also be including display unit etc., and details are not described herein.Specifically in the present embodiment In, the processor 401 in the network equipment can be corresponding by the process of one or more application program according to following instruction Executable file be loaded into memory 402, and the application program being stored in memory 402 is run by processor 401, It is as follows to realize various functions:

Multiple medical image samples are obtained, are classified to the lesion characteristics in the medical image sample, are tied according to classification Fruit constructs the promotion tree-model of the medical image sample, obtains the first classifier group, then using the first classifier group from this The lesion characteristics for meeting the first preset condition are filtered out in lesion characteristics, are obtained target lesion feature set, then, are utilized the target Lesion characteristics collection is trained preset second classifier, the second classifier after being trained, then, based on after the training the Two classifiers carry out the detection of target lesion feature to medical image to be sorted, obtain classification results.

The specific implementation of above each operation can be found in the embodiment of front, and details are not described herein.

It will appreciated by the skilled person that all or part of the steps in the various methods of above-described embodiment can be with It is completed by instructing, or relevant hardware is controlled by instruction to complete, which can store computer-readable deposits in one In storage media, and is loaded and executed by processor.

For this purpose, the embodiment of the present invention also provides a kind of storage medium, wherein being stored with a plurality of instruction, which can be located Reason device is loaded, to execute the step in the classification method of any medical image provided by the embodiment of the present invention.For example, The instruction can execute following steps:

Optionally, which can include that the medicine sample image training of target object is formed by multiple. After being specifically trained by other equipment, it is supplied to the sorter of the medical image, alternatively, can also be by the medicine shadow The sorter of picture is voluntarily trained；I.e. processor 401 can also run the application program being stored in memory 402, from And realize following functions:

Multiple medicine sample images comprising target object are acquired as training dataset, according to the training dataset to pre- If disaggregated model be trained, obtain disaggregated model.

Wherein, which may include: read-only memory (Read Only Memory, ROM), random access memory Body (Random Access Memory, RAM), disk or CD etc..

By the instruction stored in the storage medium, any medicine shadow provided by the embodiment of the present invention can be executed Step in the classification method of picture, it is thereby achieved that the classification side of any medical image provided by the embodiment of the present invention Beneficial effect achieved by method is detailed in the embodiment of front, and details are not described herein.

Classification method, device and the storage medium for being provided for the embodiments of the invention a kind of medical image above carry out It is discussed in detail, used herein a specific example illustrates the principle and implementation of the invention, above embodiments Illustrate to be merely used to help understand method and its core concept of the invention；Meanwhile for those skilled in the art, according to this The thought of invention, there will be changes in the specific implementation manner and application range, is to sum up somebody's turn to do, the content of the present specification should not manage Solution is limitation of the present invention.

Claims

1. a kind of classification method of medical image characterized by comprising

Obtain multiple medical image samples；

Classify to the lesion characteristics in the medical image sample, constructs the medical image sample according to classification results Promotion tree-model, obtain the first classifier group；

The lesion characteristics for meeting the first preset condition are filtered out from the lesion characteristics using the first classifier group, are obtained Target lesion feature set；

Preset second classifier is trained using the target lesion feature set, the second classifier after being trained；

The detection of target lesion feature is carried out to medical image to be sorted based on the second classifier after the training, obtains classification knot Fruit.

2. the method according to claim 1, wherein the lesion characteristics in the medical image sample into Row classification, the promotion tree-model of the medical image sample is constructed according to classification results, obtains the first classifier group, comprising:

It is partitioned into lesion region sample in the medical image sample, it is special to the lesion region sample extraction multidimensional lesion Sign；

Classify to the multidimensional lesion characteristics, the boosted tree mould of the medical image sample is constructed according to classification results Type obtains the first classifier group.

3. according to the method described in claim 2, it is characterized in that, the first classifier group includes multiple base classifiers, institute It states and classifies to the multidimensional lesion characteristics, the promotion tree-model of the medical image sample is constructed according to classification results, Obtain the first classifier group, comprising:

Multiple lesion characteristics are randomly choosed from the multidimensional lesion characteristics using each base classifier；

The multiple lesion characteristics are classified, to construct the Taxonomy and distribution mould of multiple medical image samples Type；

Multiple Taxonomy and distribution models are combined, the promotion tree-model of the medical image sample is constructed, obtains the first classifier Group.

4. according to the method described in claim 3, it is characterized in that, described special from the lesion using the first classifier group The lesion characteristics for meeting the first preset condition are filtered out in sign, obtain target lesion feature set, comprising:

Cross validation is carried out to multiple base classifiers in the first classifier group, obtains the accuracy rate of each base classifier；

The multiple base classifiers for selecting the accuracy rate to be greater than preset threshold are combined, and obtain object classifiers；

Target lesion feature set is filtered out based on the object classifiers.

5. according to the method described in claim 4, it is characterized in that, described filter out target lesion based on the object classifiers Feature set, comprising:

It counts base classifier accuracy rate in the cross validation and is greater than used lesion characteristics when preset threshold, obtain lesion spy The frequency of use of sign；

6. according to the method described in claim 2, it is characterized in that, described be partitioned into diseased region in the medical image sample Domain sample, to the lesion region sample extraction multidimensional lesion characteristics, comprising:

The tissue regions by lesion region and nearly lesion are filtered out in the medical image sample；

Tissue regions by the lesion region and nearly lesion are extended, be expanded rear region；

The extension rear region is split, lesion region sample is obtained；

Using medical image feature extraction packet to the lesion region sample extraction multidimensional lesion characteristics.

7. method according to any one of claims 1 to 6, which is characterized in that the multiple medical image sample to be divided into Medical image training sample and medical image verify sample, described to be classified using the target lesion feature set to preset second Device is trained, the second classifier after being trained, comprising:

Default second classifier is constructed using the target lesion feature set, using the medical image training sample to described pre- If the second classifier is trained, the second classifier is obtained；

The accuracy of second classifier is verified using medical image verifying sample, if verification result meets second in advance If condition, then the second classifier after being trained.

8. method according to any one of claims 1 to 6, which is characterized in that described based on the second classification after the training Device carries out the detection of target lesion feature to medical image to be sorted, obtains classification results, comprising:

Obtain medical image to be sorted；

The detection of target lesion feature is carried out to the medical image using the second classifier after the training；

If testing result indicates the medical image, there are target lesion features, it is determined that there are diseased regions for the medical image Domain.

9. a kind of sorter of medical image characterized by comprising

Acquiring unit, for obtaining multiple medical image samples；

Construction unit constructs institute according to classification results for classifying to the lesion characteristics in the medical image sample The promotion tree-model for stating medical image sample obtains the first classifier group；

Screening unit meets the first preset condition for filtering out from the lesion characteristics using the first classifier group Lesion characteristics obtain target lesion feature set；

Training unit, for being trained using the target lesion feature set to preset second classifier, after being trained Second classifier；

Detection unit, for carrying out the inspection of target lesion feature to medical image to be sorted based on the second classifier after the training It surveys, obtains classification results.

10. device according to claim 9, which is characterized in that the construction unit, comprising:

Subelement is extracted, for being partitioned into lesion region sample in the medical image sample, to the lesion region sample Extract multidimensional lesion characteristics；

11. device according to claim 10, which is characterized in that

The building subelement, specifically for randomly choosing multiple diseases from the multidimensional lesion characteristics using each base classifier Become feature；The multiple lesion characteristics are classified, to construct the Taxonomy and distribution of multiple medical image samples Model；Multiple Taxonomy and distribution models are combined, the promotion tree-model of the medical image sample is constructed, obtains the first classifier Group.

12. device according to claim 10, which is characterized in that

The extraction subelement, specifically for filtering out the tissue by lesion region and nearly lesion in the medical image sample Region；Tissue regions by the lesion region and nearly lesion are extended, be expanded rear region；To the extension back zone Domain is split, and obtains lesion region sample；Using medical image feature extraction packet to the lesion region sample extraction multidimensional Lesion characteristics.

13. according to the described in any item devices of claim 9 to 12, which is characterized in that described by the multiple medical image sample Originally it is divided into medical image training sample and medical image verifying sample, the training unit, comprising:

Training subelement, for constructing default second classifier using the target lesion feature set, using the medical image Training sample is trained default second classifier, obtains the second classifier；

Subelement is verified, for verifying using medical image verifying sample to the accuracy of second classifier, if testing It demonstrate,proves result and meets the second preset condition, then the second classifier after being trained.

14. according to the described in any item devices of claim 9 to 12, which is characterized in that the detection unit, comprising:

Subelement is obtained, for obtaining medical image to be sorted；

15. a kind of storage medium, which is characterized in that the storage medium is stored with a plurality of instruction, and described instruction is suitable for processor It is loaded, the step in the classification method of 1 to 8 described in any item medical images is required with perform claim.